Image registration using intrinsic features is an important task in image-guided pre-surgery planning, surgery and post-surgery analysis. Surface-based registration is one of the most commonly used techniques for structural registration (head registration, bone registration, etc). It does not require the use of external markers and is suitable for cross-modality registration of computed tomography (CT) and magnetic resonance (MR) images. However, this technique requires the precise extraction of the corresponding surfaces and also a computationally expensive optimisation routine to compute the transformation relating the reference and current surfaces. In this article, we describe a fast and near-automatic procedure that we developed for accurately registering the same bone surfaces from two CT images.

An iterative adaptive thresholding procedure for semi-automatic segmentation of bone in CT images was used to classify each pixel of the CT image into either bone or non-bone (Figure 1). A pixel is said to be on the boundary if at least one of its connected neighbors belongs to the non-bone class. Each boundary pixel is re-classified by comparing its intensity to a local optimal threshold which is computed using a small window centered on the pixel. This process is repeated until convergence. The segmentation results are shown in Figure 1.

Figure 1: Registration process: (a) Original CT dataset, (b) bone region after iterative adaptive thresholding and (c) extracted bone surface.

The type of surface modeling will directly impact the speed of the registration. The registration process will compute the optimal transformation between the two datasets by comparing the surface models extracted from these datasets. Thus, it is essential for the surface model to be computational efficient.

We used a neural network (NN) to model a function that measures the distance of any spatial point from the reference surface. This function varies smoothly across surfaces; points inside the object have a negative distance while points outside the surface have positive distance, and hence the object's surface is defined implicitly as zero distance. A neural network was used to model the surface of the bone as it is able to able to perform non-linear modelling (good for complex surfaces), has extremely low computational requirements (feed-forward NN), and can be implemented efficiently in hardware (e.g. field-programmable gate arrays). Furthermore, the intensive training required by neural network modelling is done prior to the registration. Thus, during the image guided operation, the registration only uses the feed-forward evaluation of the neural network, which is extremely fast.

The final component in the surface based registration is to find the transformation that gives the lowest cost (Figure 1 (c)). We use a subspace trust region method based on the interior-reflective Newton method. An excised section of a human vertebrae body was scanned using CT in two different orientations. The in-plane pixel size was approximately 0.25 mm. Reconstructions were done every 0.5 mm in the through-plane direction. The results show that our registration system has sub-pixel accuracy. The time taken for the registration process was less than 1 minute. We foresee that this fast and accurate registration technique will play a major role in future systems for image-guided surgery.

This research was done in collaboration with Assoc Prof SH Ong, Prof SH Teoh, Mr Chui Chee Kong and graduate student, Ge Yi.